Re: [AArch64] Handle HFAs of float16 types properly

[James Greenhalgh <james dot greenhalgh at arm dot com>]

On Fri, Aug 05, 2016 at 11:00:39AM +0100, Yao Qi wrote:
> On Tue, Jul 26, 2016 at 2:55 PM, James Greenhalgh
> <james.greenhalgh@arm.com> wrote:
> >
> > OK? As this is an ABI break, I'm not proposing for it to go back to GCC 6,
> > though it will apply cleanly there if the maintainers support that.
> >
> 
> What do you mean by "ABI break"?  AFAICS, with this patch, it conforms to
> AAPCS.

After this patch code generated for GCC 4.9/5/6 will not be ABI
compatible with code generated for GCC 7 for HFAs of __fp16. The new
generated code will conform to AAPCS64, but the old code didn't so there has
been an ABI change between the GCC versions. We don't like doing that for
minor releases, so the patch is not really suitable for backporting.

> The subject leads me thinking about the handling of HVA of float16.

These are handled like any other vector, the code looking at HVA's doesn't
care about the inner mode of the vector just the bitsize:

  config/aarch64/aarch64.c::aapcs_vfp_sub_candidate

    case VECTOR_TYPE:
      /* Use V2SImode and V4SImode as representatives of all 64-bit
	 and 128-bit vector types.  */
      size = int_size_in_bytes (type);
      switch (size)
	{
	case 8:
	  mode = V2SImode;
	  break;
	case 16:
	  mode = V4SImode;
	  break;
	default:
	  return -1;
	}

      if (*modep == VOIDmode)
	*modep = mode;

      /* Vector modes are considered to be opaque: two vectors are
	 equivalent for the purposes of being homogeneous aggregates
	 if they are the same size.  */
      if (*modep == mode)
	return 1;

      break;

Thanks,
James